Achieving Lightweight Image Steganalysis with Content-Adaptive in Spatial Domain

Abstract

Steganography is a technology that modifies complex regions of digital images to embed secret messages for the purpose of covert communication, while steganalysis is to detect whether secret messages are hidden in a digital image or not. However, the emergence of content-adaptive steganography such as S-UNIWARD prioritizes the embedding of secret messages in areas of textural complexity of images by embedding probability map guidelines. Such ways dramatically improve the security of steganography and impede the process of image steganalysis. Most of the existing steganalysis studies are aimed at improving the network structure to enhance the detection performance of the model, without considering the generation of embedding probability maps which can guide the training of the network model, eliminate some unnecessary distractions, shorten the training time and improve the final detection accuracy simultaneously. Therefore, how to obtain embedded probability maps and use them effectively becomes an important challenge in the field of steganalysis. In this paper, to solve the above problem we propose a content-adaptive lightweight network to implement an embedded probability map combined with steganalysis. Our steganalysis model includes two parts: embedding probability maps generation module and features processing module, which is trained Separately. The generation module adopts the basic framework and modifies the model to make it more suitable for steganography. In the features processing module, we adopt a pseudo-siamese architecture to manipulate two different input images. Next, we use the attention mechanism to assign weights to channel parameters. Finally, We use a simple data augmentation method to enhance our training dataset and improve final performance. Because our proposed model incorporates embedded probability maps as guidelines, experiments show that our proposed CNet has faster convergence speed, higher detection accuracy, and better robustness compared to networks such as Yedroudj-Net, SRNet, and Zhu-Net in the spatial domain.

Supported by Foundation item: National Key Research and Development Program of China (2018YFB1003205); National Natural Science Foundation of China (U1836110, U1836208); by the Jiangsu Basic Research Programs-Natural Science Foundation under grant numbers BK20200039.